Research On Object Detection Algorithm With Object Region Estimation

Object detection is one of the important topics in the field of computer vision and has widely used in many intelligent systems.This paper proposes a high-performance object detection algorithm based on object region estimation,which has achieved good detection results in infrared small target and area target detection tasks.To solve the problem of high false alarm of infrared small target detection under strong clutter,this paper first proposes an infrared small target detection algorithm based on target region estimation.The algorithm uses a fully convolutional network to estimate the region of the small target.To reduce the number of false targets in region estimation results,the algorithm proposes a target confidence network,which adopts SENet structure to fuse image features before and after background suppression to determine the authenticity of the target.Finally,the algorithm achieves a detection rate of 96% on the test-set while maintaining a low false alarm number.Aiming at the selection of positive and negative samples in area target detection,this paper overcomes the shortcomings of the Anchor matching mechanism by estimating the center area of the target.The region estimation algorithm first uses the ellipse Gaussian diffusion function to describe the target center region,and uses estimation model based on U-Net structure.In the estimation model,this paper introduces a self-attention mechanism to enhance the expressive ability of features.Aiming at the problem that the receptive field of the output layer is difficult to express larger targets,this paper builds a correlation feature pyramid network,which improves the accuracy of the target center area estimation.Finally,a target detection algorithm based on the target center region estimation model is proposed.For the time-consuming problem of two-stage detection algorithm,this paper proposes the regional feature pooling algorithm to extract the target feature,which accelerates the feature extraction and processing speed.Aiming at the diversity of target feature scales,a cascade structure based on astrous convolution is introduced to improve the receptive field of the model.Aiming at the problem that the prior information of Anchor mechanism is not flexible enough,a cascade regression of the deviation of the wide and high branches is proposed to obtain a more accurate size.To predict the target category,a method for weighting the output of category cascades based on the target size is proposed.The results show that the regional feature pooling algorithm and cascade structure greatly improve the detection performance.Among the mainstream methods compared,the algorithm proposed in this paper achieves the best detection accuracy and maintain a high processing frame rate.